Transcription
Hello everyone. My name is Kostya. Today we will be testing this model called Step 35 Flash. The model is extremely interesting. It has 256,000 context, and right now it is completely free. You can use it for free on Open Router. That is, you need to create an account here, top it up with 5 dollars using crypto, for example, and use all the free models that are on Open Router. There are quite a few of them. Let's talk about the model itself. The claims are, of course, very interesting. So, here is its rating, yes, 81. It turns out that it is even a little better than Clot Opus 4.5 and Gemini 3 Pro. And also better than KI K25. And slightly worse than GPT 5.2 X. Interesting. Interesting. The claim is quite good. Let's take a quick look at the benchmarks. Reasoning for it. Uh-huh. So it can even think stronger than GLM 4.7. The problem with GLM 4.7 is that it is very slow, while, for example, Step 3.5 Flash is very fast. This is its undeniable advantage. If it, uh, shows itself better than GLM 4.7, and at the same time is very fast, then it will be a huge replacement for GLM. In coding, it is slightly better than GLM by about one or half a point. Uh-huh. Terminal benchmark, lifecoding benchmark. That is, lifecoding. Why are there two values here? What does the second one mean? Understood. And what? Nothing is understood. Okay. So it is at the level of GPT 5.2 here. Yes, and it hasn't gone far from here. In general, it should be strong in coding, which we will check today. I will use Kilocode for all of this. It was easiest for me to connect. Of course, I could have used Open Code, but visually Kilocode just looks more interesting. So, I have an empty folder open here. I connected Step 3 Flash here. Everything connects easily. You need to install Visual Studio Code or Antigravity, or something else. And in extensions, you need to type Kilocode, install it. Accordingly, you will see this window, like the one I have. So, here is Kilocode. And here I can do the following. So I will need to go into settings, create a new provider. I will write here, for example, Open Router. Yes, I will not save it, of course, you will just need to create a profile. Then, from the list of providers, you will find Open Router here. This is what it looks like. Enter your API key here. You can get it on the Open Router website in the Keys section. And here you will need to find that very Step 3.5 Flash model. Make sure it is typed here. Then you will need to set it in each mode. So, like this, find Open Router and set it here in each mode as well. It is already set in all modes for me. I will test this neural network in three different types of projects. This is a landing page, this is some kind of web application, and this is a mobile application. Let's start with the landing page. So, for this, I will go to Anti, or rather, to Codewhisperer. Here I wrote this very detailed technical specification for the neural network, which will create this project in a file. In the root of the project, in English. This website is dedicated to providing services in the field of website creation. Come up with all the blocks yourself. In general, I wrote what kind of technical specification I need to create for our neural network. That is, using GPT 5.2 Codewhisperer, I will ask it to create a technical specification, which it will write, so that based on this technical specification, our dear Step 3.5 Flash will create a cool landing page for us. So, our GPT said that it completed the task. Now we will go back to Antigravity. And we have this file here, which says the technical specification, what needs to be done. Now we will write to this person that we need. Create, just create a website, create, rather, a landing page according to the technical specification from the file, from the file. And we will drag it simply not like this, but like this. Technical specification and send. Let's see what happens. So, it has started. They say that this neural network is very fast. Let's see how long it takes. It really did it quite quickly. That is, it took literally 5 minutes. Now let's see what it did for us in 5 minutes. I am opening the project. Well, let's evaluate what's here. Does it look like Star Wars? No, not at all. Does it look like something cosmic? Well, yes, it does. Is it beautiful? No. Are there any cool blocks here? Is there any animation? No. So what is this? And you still write to me that you can't understand what the neural network is capable of from a landing page. You can. This complete mess it made. Well, it's very weak. Compared to Gemini, compared to Mika 2 with, it's just nothing. Although usually neural networks do a great job with these GPT technical specifications, they create designs and blocks, and so on. So, let's create some web project. I will open a new window here and type this. Create, create a new folder web and install next.js in it with at least these simple tasks it can handle. But it was claimed that it is very strong, cool, very weak in design, but we are not judging by design. Maybe it is stronger in other tasks, for example, technically complex ones. Okay, let it create the project for now. We have another task. We have Codewhisperer. I will ask it to create a very detailed technical specification for a web project, a to-do list, personal efficiency, habits with a drag-and-drop effect and a timer in a modern style. Let GPT create a new technical specification for us. Let's see if this one is creating the project. Yes. Create next web. Yes. St. Here is our project. Everything is ready. We also have the technical specification. Now we have a new task. So, in the folder, folder web, create a project based on this technical specification. Let's send it again to create a project based on technical specification number two and see what it does for us this time. This took quite a long time. I saw a lot of errors pouring in, pouring in, pouring in, pouring in. It kept solving them, solving them, this way and that way. I'm even afraid to look at what's on localhost 3000, but we'll look anyway. Well, and what is this? So, well, for example, the layout, I see, is broken. Okay. And how do I add a new task? Uh, and what's here? Nothing is clickable, nothing works. Well, as expected. And who will tell me that it shouldn't be like this? You can't create a web application from one prompt. Why do other neural networks create applications from one prompt, but this one, this one, this one, Step Flash can't? Why can others, but this one can't? So, it's not very good. And, well, I don't think there's even any point in making a Flutter application, because, well, the neural network is showing itself to be not very good. Maybe it shows itself differently for you. If it shows itself differently for you, please test it, show, well, tell me about your results in the comments. Maybe I have some specific conditions or GPT wrote a bad prompt, so the neural network didn't cope. Although, if I now throw the same prompt into KIC 2, let me do it right now and throw this same prompt into KIC 22, and we'll see what it does and what kind of web application it will assemble for us. Just to see the difference in the fact that it claims that it is supposedly there, that it has Mika 2 with, it's better and so on. So. Yes, it says that there are only 1,000 parameters and 8.2, yes, score. And it has 196 parameters, 1,000, and it has an 81 score, like it's much cooler. Let's, let's see, cooler or not cooler. Well, let's conduct an experiment. I'm just curious myself. So, I'm going to Antigravity. We will still use Kilocode for the sake of experiment. Let's ask. Now I will select Mika 2 with here. Mi, I think I have money here. Yes, yes, I have money here. Then I will select the default profile everywhere. Default, default, default. And here default. Everything, I have the default profile everywhere. And in the default profile, I will just change it now to Mika 2 with. Here it is, Mika 2 with. And we will now see here, only in the orchestrator, not default. And we will now see how well Mika 2 with copes with this same task, and whether it copes from one prompt, which is the most important thing. I think it will cope. So, I am creating a new task, a new task, and I am writing it approximately the same. So. And I will even copy C. So, we don't need this. Everything, I'm entering the same prompt into the folders. So, only in the folder. Uh-huh. Create. So, okay. Then we will ask it to create a new folder. Create. Create folder web 2 and install next.js there. Let's see. Now it will create Next there. And it will probably do it even faster. And then we will enter the next prompt about what it should do. In principle, I will launch it immediately as the second one. So, as soon as it completes its task, we will again open exactly the same project and see who is better. Here, I remind you, nothing works at all. Well, that is, pages switch. Oh, something, something still works here, and it's also terrible. That is, even. Oh, more errors are pouring in immediately. These are errors. It's not that something stopped working today, errors are just pouring in. M. Yes. Okay. Let's see what the task 2 with will create then. So, the project on Mika 2 with has been created. I have already looked at what it created. It created a spaceship. And you can add different tasks here. I will add randomly, that is, just some data. Well, I'll write the tag correctly. So, what is this, well, editing, by the way, doesn't work. Okay, let's assume, but the fact remains that applications are created from the first try. Yes, it can be, and it should be, and can be improved further. But the point is, why are projects created, that the focus timer works, that statistics will also be collected, I am sure. For example, that I can also, for example, do some task, bang, if it's completed, then it moves to completed. It's not difficult for a neural network to create some project. The thing is that, unfortunately, Step Flash is not, uh, not a neural network that you can work with in the long term, although it shows, well, once again I am convinced that all these so-called benchmarks are worth nothing, that you need to test everything yourself, on your own projects, on your own tasks, and draw your own conclusions about how a particular neural network works. For example, many people don't like GLM, although with GLM, in many ways, my relationship has been quite good, but at the same time, some Step 3.5 Flash came out and in the comments under previous videos they praised it, saying: "It's cool, it works fast." But in fact, it turned out not very good. Not very good at all. And in general, I wanted to talk about these Chinese neural networks. If you have absolutely no budget, if you are willing to spend 0 rubles or 3 dollars, even if you feel sorry for GLM or you only have 3 dollars and you are willing to spend it only on GLM, then okay, no questions. Like, 0% judgment, 100% understanding, but for those who do have money and are willing to buy a neural network, I still advise you to go for the top three. This is Gemini, this is Clotd, this is ChatGPT, well, GPT Codewhisperer, I mean. That's it, I wouldn't advise using anything else. Well, Mika 2 with can also be used because it creates good designs. Specifically from the design point of view, it's not bad, just like GI in many ways does well with design, it creates interfaces. And even then, I would argue which one is better. In terms of logic, of course, Clot OPUS does logic well. It's a great executor, it writes code. In my opinion, it plans the architecture much better, Clot OPUS plans the architecture much better. So, in conclusion, I started with a review and ended with a comparison. And in the comparison, Mika 2 wins against Step 3.5 Flash. Thank you all for watching. And here's a rather peculiar video. Write your comments, write your opinion about all of this, what you think about all of this, which neural networks you develop on. Be sure to write in the comments, tell me what you use, what projects you do at all. And I say goodbye to you. See you soon.